VoxForge
VoxForge
Hello
I'm working on discriminative training. However i found some
problems in implementation under HTK. My problem is when i try to
generate the Word Lattice using HDecode , I always get the same
following errors:
HDecode -A -H hmms1/hmm15/models -S
train_lattice.scp -t 220.0 220.0 -C config.hdecode -i
wlat.den/recout.mlf -w trainbg -o M -z lat -l wlat.den -X lat -p 0.0 -s
5.0 dict.hdecode tiedlist
ERROR [+9999] HLVNet: no model label for phone (td-f+ao)
FATAL ERROR - Terminating program HDecode
And when I execute the command once more i found other error:
ERROR [+9999] HLVNet: no model label for phone (td-s+eh)
.
.
ERROR [+9999] HLVNet: no model label for phone (td-n+ay)
and so on in loop
I
found the instruction about the garbage model but i can't solve the
problem since each phoneme above belongs to the word pronunciation.
On
other hand there are no word who has as pronunciation as the triphones
above. And latter are neither present in HMMdefintion nor tiedlist.
i can't understand from where HDecode got these triphones.
it seems that the phone "td" are add to each Pronunciation word in the dictionary whereas it should not be added.
dictionary look like:
EIGHT ey td
FIVE f ay v
FOUR f ao
FOUR f ao r
NINE n ay n
OH ow
ONE w ah n
SENT-END [] sil
SENT-START [] sil
SEVEN s eh v ax n
SIX s ih k s
THREE th r iy
TWO t uw
ZERO z ia r ow
The word-level training transcripts with one sentence look:
SENT-START ONE SENT-END
SENT-START TWO SENT-END
SENT-START THREE SENT-END
SENT-START FOUR SENT-END
SENT-START FIVE SENT-END
SENT-START SIX SENT-END
SENT-START SEVEN SENT-END
SENT-START EIGHT SENT-END
SENT-START NINE SENT-END
SENT-START ZERO SENT-END
SENT-START OH SENT-END
SENT-START and SENT-END are used as start/ endnode.
I will be very grateful if any help me solve this problem.
--- (Edited on 7/2/2012 1:17 pm [GMT-0500] by raf23) ---
> it seems that the phone "td" are add to each Pronunciation word in the dictionary whereas it should not be added.
HDecode is a large-vocabulary decoder. It considers all possible combinations. So the triphones with td- are perfectly natural.
If you want to run MMI training with a small vocabulary you can create lattices with HVite, not with HDecode. It is possible in a latest HTK version.
--- (Edited on 7/3/2012 23:27 [GMT+0400] by nsh) ---
I have the same problem! any solution please
--- (Edited on 6/8/2014 6:11 am [GMT-0500] by Visitor) ---
The solution is to make the"fulllist" in making tied state triphone step, contains ALL possible mono, bi and tri gram phone using full_list.prl script
Moreover, hdecode works with LM in ARPA format, this LM can be generated using SRILM
--- (Edited on 6/8/2014 11:38 pm [GMT-0500] by Visitor) ---
I generated all mono, bi and tri phones using full_list.prl and I now have problem running HHEd, as follows:
ERROR [+2662] FindProtoModel: no proto for bo-cm+sil in hSet
FATAL ERROR - Terminating program HHEd
When I remove all of the tri phones that do not have models from fulllist (since I do not have too much training data), training steps run properly, but I end up having the following error when I run HDecode:
ERROR [+9999] HLVNet: no model label for phone (sil-am+bd)
when I run it again, it produces:
ERROR [+9999] HLVNet: no model label for phone (sil-cd+ce)
which is basically a different triphone.
Any solutions to this problem
--- (Edited on 8/14/2014 9:23 pm [GMT-0500] by wamageed) ---